The first-of-its-kind architecture gives up to a 7x speedup over traditional training methods.